The global ERP market surpassed $59 billion in 2023, yet these heavyweight platforms remain little more than expensive digital archives. Researchers from the KTH Royal Institute of Technology—Zihao Liu, Tianyu Wang, Xi Vincent Wang, and Lihui Wang—point out the obvious: modern ERPs excel at recording historical facts but freeze when forced to make operational decisions. When a supplier delays a shipment or demand spikes, the system waits passively for a human to decipher the situation and manually update the records. This ceiling is dictated by the limitations of traditional automation: rigid algorithms cannot handle exceptions, and monolithic AI assistants degrade rapidly when tasked with simultaneously coordinating procurement, logistics, and finance.
The Architecture of Role-Based Autonomy
To overcome the intellectual burnout of single models, the KTH team proposes the concept of Agentic ERP. This is not just another "smart" chatbot bolted onto a database, but a decentralized system where operational tasks are distributed among LLM agents with clearly defined roles. This specialization radically reduces the complexity of tool selection at each step: every agent knows its maneuver. At the heart of the system is a graph-based orchestrator utilizing a Planner–Executor–Reflector–Responder framework. It separates planning from execution and evaluation, vetting every action against external criteria and "sprint contracts" before it hits the production environment.
This research proves that a network of role-based LLM agents under human supervision can transform ERP from a passive transaction logger into an active decision-making hub.
The system interprets ERP operations as a sequential decision-making task under strict constraints. This allows agents to confidently navigate hundreds of screen forms and transaction codes that typically intimidate even experienced users. Instead of forcing a human to wander through a labyrinth of fields to reorder parts, the agent layer uses the ERP as an execution backend for its autonomous workflows.
Dynamic Reasoning vs. Rigid Algorithms
Traditional automation via RPA (Robotic Process Automation) breaks down at the first sign of an anomaly because it exists in a world of "if-then" logic. Agentic ERP replaces this brittle approach with dynamic reasoning. In a year-long simulation, the multi-agent approach outperformed both RPA and classical manual methods. The data is impressive: the system maintained 365 days of operation with zero stockouts, while rigid algorithms accumulated hundreds of failures under identical demand fluctuations. LLMs effectively manage the cognitive load required to interpret anomalies—such as non-standard customer requests—that typically cause standard automation to crash.
The multi-agent method significantly outperforms baseline solutions, ensuring a year of operation without stockouts while RPA scripts drown in errors.
Despite the push for autonomy, handing the steering wheel entirely to algorithms is corporate suicide. Therefore, the architecture includes a Risk-Tiered Human-in-the-Loop concept. This safety valve determines the risk level of a task: if the stakes are low, the agent acts independently; if a critical failure looms, it calls for a human. This approach transforms the ERP from a "data graveyard" into an expert system capable of finally reducing the "cognitive tax" on specialists, sparing them from endless manual clicking in hostile interfaces. Moving toward autonomous management requires abandoning monoliths in favor of role-based structures. While integrating into a landscape of legacy systems remains a challenge, the path to an ERP that manages the future rather than just remembering the past is now clearly visible.